Graphics Performance Optimisation John Spitzer Director of European Developer Technology
Overview Understand the stages of the graphics pipeline Cherchez la bottleneck Once found, either eliminate or balance
Simplified Graphics Pipeline CPU Geometry Storage Geometry Processor Rasterizer Fragment Processor Frame buffer Texture Storage + Filtering Vertices Pixels
Possible Pipeline Bottlenecks CPU transfer transform raster texture fragment frame buffer CPU Geometry Storage Geometry Processor Rasterizer Fragment Processor Frame buffer CPU/Bus Bound Vertex Bound Texture Storage + Filtering Pixel Bound
Battle Plan for Better Performance Locate the bottleneck(s) Eliminate the bottleneck (if possible) Decrease workload of the bottlenecked stage Otherwise, make it look better Balance pipeline by increasing workload of the non-bottlenecked stages
Bottleneck Identification Run App Vary FB FPS varies? Yes FB limited No Vary texture size/filtering Vary resolution Vary vertex instructions FPS varies? No FPS varies? No FPS varies? No Yes Yes Yes Texture limited Vary fragment instructions Transform limited Yes FPS varies? No Fragment limited Raster limited Vary vertex size/ AGP rate FPS varies? No Yes Transfer limited CPU limited
CPU Bottlenecks CPU transfer transform raster texture fragment frame buffer CPU Geometry Storage Geometry Processor Rasterizer Fragment Processor Frame buffer CPU/Bus Bound Vertex Bound Texture Storage + Filtering Pixel Bound
CPU Bottlenecks Application limited (most games are in some way) Driver or API limited too many state changes (bad batching) using non-accelerated paths Use VTune (Intel performance analyzer) caveat: truly GPU-limited games hard to distinguish from pathological use of API
Consolidate Small Batches Each vertex buffer/array preferably has thousands of vertices or more Draw as many triangles per call as possible ~50K DIPs/s COMPLETELY saturate 1.5GHz Pentium 4 50fps means 1,000 DIPs/frame! Up to you whether drawing 1K tri/frame or 1M tri/frame
Batch Consolidation Strategies Use degenerate triangles to join strips together Hardware culls zero-area triangles very quickly Use texture pages Use a vertex shader to batch instanced geometry VS2.0 and VP30 have 256 constant 4D vectors
Geometry Transfer Bottlenecks CPU transfer transform raster texture fragment frame buffer CPU Geometry Storage Geometry Processor Rasterizer Fragment Processor Frame buffer CPU/Bus Bound Vertex Bound Texture Storage + Filtering Pixel Bound
Geometry Transfer Bottlenecks Vertex data problems size issues (just under or over 32 bytes) non-native types (e.g. double, packed byte normals) Using the wrong API calls Immediate mode, non-accelerated vertex arrays Non-indexed primitives (e.g. gldrawarrays, DrawPrimitive) AGP misconfigured or aperture set too small
Optimising Geometry Transfer: OpenGL Static geometry display lists okay, but ARB_vertex_buffer_object is better Dynamic geometry - use ARB_vertex_buffer_object vertex size ideally multiples of 32 bytes (compress or pad) access vertices in sequential (cache friendly) pattern always use indexed primitives (i.e. gldrawelements) 16 bit indices can be faster than 32 bit
Optimising Geometry Transfer: Direct3D Static geometry: Create a write-only vertex buffer and only write to it once Dynamic geometry: Create a dynamic vertex buffer Lock with DISCARD at start of frame Then append with NOOVERWRITE until full Use NOOVERWRITE more often than DISCARD Each DISCARD takes either more time or more memory So NOOVERWRITE should be most common Never use no flags
Geometry Transform Bottlenecks CPU transfer transform raster texture fragment frame buffer CPU Geometry Storage Geometry Processor Rasterizer Fragment Processor Frame buffer CPU/Bus Bound Vertex Bound Texture Storage + Filtering Pixel Bound
Geometry Transform Bottlenecks Too many vertices Too much computation per vertex Vertex cache inefficiency
Too Many Vertices Favor triangle strips/fans over lists (fewer vertices) Use levels of detail (but beware of CPU overhead) Use bump maps to fake geometric detail
Too Much Vertex Computation: Fixed Function Avoid superflous work >3 lights (saturation occurs quickly) local lights/viewer, unless really necessary unused texgen or non-identity texture matrices Consider commuting to vertex program if (and only if) good shortcut exists example: texture matrix only needs to be 2x2 not recommended for optimizing fixed function lighting
Too Much Vertex Computation: Vertex Programs Move per-object calculations to CPU, save results as constants Leverage full spectrum of instruction set (LIT, DST, SIN,...) Leverage swizzle and mask operators to minimize MOVs Consider using shader levels of detail
Vertex Cache Inefficiency Always use indexed primitives on high-poly models Re-order vertices to be sequential in use (e.g. NVTriStrip) Favor triangle fans/strips over lists
Rasterization Bottlenecks CPU transfer transform raster texture fragment frame buffer CPU Geometry Storage Geometry Processor Rasterizer Fragment Processor Frame buffer CPU/Bus Bound Vertex Bound Texture Storage + Filtering Pixel Bound
Rasterization Rarely the bottleneck (exception: stencil shadow volumes) Speed influenced primarily by size of triangles Also, by number of vertex attributes to be interpolated Be sure to maximize depth culling efficiency
Maximize Depth Culling Efficiency Always clear depth at the beginning of each frame clear with stencil, if stencil buffer exists feel free to combine with color clear, if applicable Coarsely sort objects front to back Don t switch the direction of the depth test mid-frame Constrain near and far planes to geometry visible in frame Use scissor to minimize superfluous fragment generation for stencil shadow volumes Avoid polygon offset unless you really need it NVIDIA advice use depth bounds test for stencil shadow volumes
Texture Bottlenecks CPU transfer transform raster texture fragment frame buffer CPU Geometry Storage Geometry Processor Rasterizer Fragment Processor Frame buffer CPU/Bus Bound Vertex Bound Texture Storage + Filtering Pixel Bound
Texture Bottlenecks Running out of texture memory Poor texture cache utilization Excessive texture filtering
Conserving Texture Memory Texture resolutions should be only as big as needed Avoid expensive internal formats New GPUs allow floating point 4xfp16 and 4xfp32 formats Compress textures: Collapse monochrome channels into alpha Use 16-bit color depth when possible (environment maps and shadow maps) Use DXT compression
Poor Texture Cache Utilization Localize texture accesses beware of dependent texturing beware of non-power of 2 textures ALWAYS use mipmapping use trilinear/aniso only when necessary (more later!) Avoid negative LOD bias to sharpen texture caches are tuned for standard LODs sharpening usually causes aliasing in the distance opt for anisotropic filtering over sharpening
Excessive Texture Filtering Use trilinear filtering only when needed trilinear filtering can cut fillrate in half typically, only diffuse maps truly benefit light maps are too low resolution to benefit environment maps are distorted anyway Similarly use anisotropic filtering judiciously even more expensive than trilinear not useful for environment maps (again, distortion)
Fragment Bottlenecks CPU transfer transform raster texture fragment frame buffer CPU Geometry Storage Geometry Processor Rasterizer Fragment Processor Frame buffer CPU/Bus Bound Vertex Bound Texture Storage + Filtering Pixel Bound
Fragment Bottlenecks Too many fragments Too much computation per fragment Unnecessary fragment operations
Too Many Fragments Follow prior advice for maximizing depth culling efficiency Consider using a depth-only first pass shade only the visible fragments in subsequent pass(es) improve fragment throughput at the expense of additional vertex burden (only use for frames employing complex shaders)
Too Much Fragment Computation Use a mix of texture and math instructions (they often run in parallel) Move constant per-triangle calculations to vertex program, send data as texture coordinates Do similar with values that can be linear interpolated (e.g. fresnel) Consider using shader levels of detail Use lowest pixel shader version you can
GeForceFX-specific Optimisations Use even numbers of texture instructions Use even numbers of blending (math) instructions Use normalization cubemaps to efficiently normalize vectors Leverage full spectrum of instruction set (LIT, DST, SIN,...) Leverage swizzle and mask operators to minimize MOVs Minimize temporary storage Use 16-bit registers where applicable (most cases) Use all components in each (swizzling is free) Use ps_2_a profile in HLSL
Framebuffer Bottlenecks CPU transfer transform raster texture fragment frame buffer CPU Geometry Storage Geometry Processor Rasterizer Fragment Processor Frame buffer CPU/Bus Bound Vertex Bound Texture Storage + Filtering Pixel Bound
Minimizing Framebuffer Traffic Collapse multiple passes with longer shaders (not always a win) Turn off Z writes for transparent objects and multipass Question the use of floating point frame buffers Use 16-bit Z depth if you can get away with it Reduce number and size of render-to-texture targets Cube maps and shadow maps can be of small resolution and at 16-bit color depth and still look good Try turning cube-maps into hemisphere maps for reflections instead Can be smaller than an equivalent cube map Fewer render target switches Reuse render target textures to reduce memory footprint Do not mask off only some color channels unless really necessary
Finally... Use Occlusion Query Use occlusion query to minimize useless rendering It s cheap and easy! Examples: multi-pass rendering rough visibility determination (lens flare, portals) Caveats: need time for query to process can add fillrate overhead
Tools: NVPerfHUD Drivers now support NVPerfHUD Overlay that shows vital various statistics as the application runs Top graph shows : Number of API calls Draw*Prim*, render states, texture states, shader states Memory allocated AGP and video Bottom graph shows : GPU Idle Graphics HW not processing anything Driver Time Driver doing work (state and resource management, shader compilation) Driver Idle Driver waiting for GPU to finish Frame Time Milliseconds per frame time
NVPerfHUD - Screenshot
Conclusion Complex, programmable GPUs have many potential bottlenecks Rarely is there but one bottleneck in a game Understand what you are bound by in various sections of the scene The skybox is probably texture limited The skinned, dot3 characters are probably transfer or transform limited Exploit imbalances to get things for free
Questions, comments, feedback? John Spitzer, spit@nvidia.com